Update README.md
Browse files
README.md
CHANGED
|
@@ -47,7 +47,7 @@ base_model:
|
|
| 47 |
|
| 48 |
<p align="justify" style="font-size: 0.92em; color: #555555; line-height: 1.5;"><b>Figure 1.</b> <i>(Left)</i> Performance on MMLU (language ability) across composable pretraining stages (LM → +MMU → +T2I). Standard MoE and structurally-isolated MoT suffer catastrophic routing collapse upon integrating continuous generative objectives. <b>Rosetta maintains a stable semantic anchor throughout all stages.</b> <i>(Right)</i> Qualitative image generation results from the Rosetta model.</p>
|
| 49 |
|
| 50 |
-
---
|
| 51 |
|
| 52 |
## 🏗️ Architecture
|
| 53 |
|
|
@@ -57,7 +57,7 @@ base_model:
|
|
| 57 |
|
| 58 |
<p align="justify" style="font-size: 0.92em; color: #555555; line-height: 1.5;"><b>Figure 2. Rosetta FFN.</b> Three mechanisms enable non-destructive modality expansion: <b>(1) Unified Attention</b> — globally shared QKV projections preserve dense cross-modal interactions. <b>(2) Composable FFN</b> — modality-specific plug-and-play experts (Text / ViT / VAE) are bridged by a single Global Shared Expert that anchors foundational knowledge. <b>(3) Conflict-Free Optimization (MAOP)</b> — surgically neutralizes destructive gradients with zero memory overhead.</p>
|
| 59 |
|
| 60 |
-
---
|
| 61 |
|
| 62 |
## 📊 Benchmarks
|
| 63 |
|
|
|
|
| 47 |
|
| 48 |
<p align="justify" style="font-size: 0.92em; color: #555555; line-height: 1.5;"><b>Figure 1.</b> <i>(Left)</i> Performance on MMLU (language ability) across composable pretraining stages (LM → +MMU → +T2I). Standard MoE and structurally-isolated MoT suffer catastrophic routing collapse upon integrating continuous generative objectives. <b>Rosetta maintains a stable semantic anchor throughout all stages.</b> <i>(Right)</i> Qualitative image generation results from the Rosetta model.</p>
|
| 49 |
|
| 50 |
+
<!-- --- -->
|
| 51 |
|
| 52 |
## 🏗️ Architecture
|
| 53 |
|
|
|
|
| 57 |
|
| 58 |
<p align="justify" style="font-size: 0.92em; color: #555555; line-height: 1.5;"><b>Figure 2. Rosetta FFN.</b> Three mechanisms enable non-destructive modality expansion: <b>(1) Unified Attention</b> — globally shared QKV projections preserve dense cross-modal interactions. <b>(2) Composable FFN</b> — modality-specific plug-and-play experts (Text / ViT / VAE) are bridged by a single Global Shared Expert that anchors foundational knowledge. <b>(3) Conflict-Free Optimization (MAOP)</b> — surgically neutralizes destructive gradients with zero memory overhead.</p>
|
| 59 |
|
| 60 |
+
<!-- --- -->
|
| 61 |
|
| 62 |
## 📊 Benchmarks
|
| 63 |
|